clone-reviewer

name	clone-reviewer
description	Use when the user wants to create a new "review-as-X" persona by cloning a reviewer's public GitHub comment history. Triggers on "clone reviewer", "create review persona", "make a reviewer bot for X", "review-clone setup", or the slash command /clone-reviewer.

Clone Reviewer — derive a persona from real review history

You are setting up a new review-clone persona. The user wants you to derive a "review-as-X" voice + ruleset from one or more GitHub reviewers' public comment history, and register a slash command /review-as-<alias> for future use.

Storage (locked)

• Persona files: ~/.claude/review-clone/<alias>/PERSONA.md
• Raw scrape: ~/.claude/review-clone/<alias>/raw/pr-<n>.json
• Snapshot: ~/.claude/review-clone/<alias>/snapshot.json
• Per-persona slash command: ~/.claude/commands/review-as-<alias>.md
• Drift log overflow: ~/.claude/review-clone/<alias>/drift.log

Conventions

<plugin> throughout this skill refers to the absolute path of the review-clone plugin directory in this Claude Code install. When invoking shell commands, the calling session resolves this from its plugin loader context (typically ~/.claude/plugins/cache/... or similar). When you see <plugin>/scripts/collect.py below, substitute the resolved absolute path.

Procedure

Step 0 — Re-clone check

Check if any personas already exist. Run:

ls ~/.claude/review-clone/ 2>/dev/null | while read d; do
  [ -f ~/.claude/review-clone/$d/PERSONA.md ] && echo "$d"
done

If the user's intended alias appears in the output, present this 3-way prompt:

A persona <alias> already exists (last scanned <last_scanned_at>).

1. Refresh — pull comments since last scan, fold into existing rules
2. Modify handles & re-extract — add/remove handles, full re-pull for added handles
3. Fork to new alias — create a separate persona under a different name

Only proceed to Step 1 after the user picks. If they pick Refresh, hand off to the review-as skill's refresh flow.

Step 1 — Conversational setup (one question at a time)

Ask each of these in turn, accepting defaults shown in brackets:

1. Alias — kebab-case, unique. Becomes the slash command name (jen → /review-as-jen).
2. GitHub handle(s) — one or more, comma-separated. Multi-handle personas aggregate across all listed handles.
3. Repo scope [defaults to the current repo's origin if any] — e.g. wayflyer/wayflyer.
4. Path filters [empty = all] — comma-separated path prefixes (frontend/, packages/).
5. Extension filters [empty = all] — comma-separated extensions (.ts, .tsx, .css, .scss).
6. Time window [6 months] — hard cap at 6 in v1.
7. Tone [copy] — copy (mimic voice) or neutral (rules in plain prose).
8. Authoring voice [yes if tone=copy, else no] — include their PR descriptions as input for an "author-as-X" sub-mode.
9. Lets-go calibration [skippable] — "Anything you've seen explicitly NOT care about?" Free text, used as anti-rules.

Step 2 — Privacy notice (non-blocking)

After answers, before scraping, print:

🔔 This will derive a persona from 's public GitHub review comments. They will not be notified. A disclaimer line stating the derived nature will be written into PERSONA.md.

Continue without confirmation — non-blocking by design.

Step 3 — Run the collector

Invoke the scraper:

python <plugin>/scripts/collect.py \
  --alias <alias> \
  --handles <handles> \
  --repo <repo> \
  --months <months> \
  --paths <paths> \
  --extensions <extensions>

Capture stdout (the JSON snapshot). Stream stderr to the user so they see progress.

Step 4 — Pre-extract gate

Read the counts field from snapshot.json. Display:

Scraped PRs, <review_comments + issue_comments> comments, <pr_descriptions> PR descriptions for <alias>.

If prs > 100 OR (review_comments + issue_comments) > 200, prompt:

This is a large corpus. Theme extraction will use significant context tokens. Proceed? [y/N]

If user declines, leave the raw data in place (they can re-run later) and stop.

Step 5 — Theme extraction (LLM-driven, you do this)

Read every ~/.claude/review-clone/<alias>/raw/pr-*.json. For each comment:

• Honor withdrawals. If a comment's reply_thread shows the author conceding (e.g. "good point, ignore my last", "Ha, I conflicted myself!"), mark that comment as withdrawn — do NOT derive a rule from it.
• Multi-handle conflict resolution. When two handles' comments contradict, pick the rule with the most citations across the corpus. Preserve losers under an also_seen: field on the rule.

Derive:

1. Rules. Cluster comments by topic. For each cluster, write a rule with: title, topic tags, severity (use the reviewer's actual phrasings to calibrate), citation URL (the strongest-anchor comment), one positive + one negative example pulled directly from the corpus.
2. Voice patterns. Openers, severity ladder (block/strong/suggest/question/non-blocking), quirks (e.g. "Needs reverted" without "to be"), what they NEVER say.
3. Lets-go. From the corpus + the user's Step 1.9 input.

Step 6 — Write PERSONA.md

Use the persona_io helpers (from scripts.persona_io import write_persona, persona_path). Frontmatter shape:

alias: <alias>
handles:
  - <handle1>
repo: <repo>
filters:
  paths: [...]
  extensions: [...]
window:
  months: <n>
  since: "<iso>"
tone: <copy|neutral>
authoring_voice: <true|false>
last_scanned_at: "<iso>"
snapshot:
  prs: <n>
  review_comments: <n>
  issue_comments: <n>
  pr_descriptions: <n>
output_default: null
drift_log: []
drift_log_archived_count: 0
disclaimer: "Derived from public GitHub review history of <handles>. Not endorsed by them."

Body sections: ## Voice & tone, ## Rules (one ### per rule, with severity/citation/examples), ## What they let go, ## Drift log.

Step 7 — Write the per-persona slash command

Render <plugin>/templates/review-as-command.md.tmpl substituting <alias> and <last_scanned_at>. Write to ~/.claude/commands/review-as-<alias>.md. Create the ~/.claude/commands/ dir if missing.

Step 8 — Final summary

Print:

✅ Persona <alias> cloned.

• rules derived from comments across

  PRs

• Slash command /review-as-<alias> ready
• Window: months · last scanned <last_scanned_at>

Try it: /review-as-<alias> to review the current branch, or /review-as-<alias> chat "would they like X?".

Constraints

• Do NOT scrape outside the user's stated window. Hard cap 6 months in v1.
• Do NOT skip the withdrawal pass — false rules from softened comments are the #1 quality problem.
• Do NOT invent rules without a citation URL. Every rule must point at a real comment.
• Do NOT fabricate symbols/APIs in examples. Quote the comment's body verbatim where it names anything.